{"id":"2d274b39-6ded-42fe-8ce5-7c9f27dcb05d","arxiv_id":"2412.05217","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"On stationary random graphs satisfying quantitative connectedness, nonlinear minimum-cost flow problems Gamma-converge under rescaling to a continuum divergence-constrained problem with a homogenised energy density from a cell formula.","lead":"This paper proves that minimum-cost flow problems on large stationary random graphs converge, after rescaling, to a continuum optimisation problem with an effective cost density. It extends periodic-graph homogenisation to non-periodic random graphs and covers nonlinear, nonconvex, multi-species costs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the claimed Gamma-convergence theorem is convincingly supported, with F3-additivity the acknowledged structural hinge and no hidden circularity found in the proof.","rationale":"The paper's central claim is Theorem 2.8, a Gamma-convergence statement for constrained minimum-cost flow functionals on stationary random graphs. I examined the proof structure: the cell formula in Definition 6.4, the subadditivity in Lemma 6.6(iii), the existence of uniform flows in Proposition 5.4, the correctors in Proposition 7.4, the upper bound in Section 9, and the lower bound in Section 10. The proof is long but the logical structure is coherent and the key estimates are stated with explicit error terms. The reader's weakest-assumption selection (F3 sigma-additivity) is spot-on: it is the one structural condition that the proof uses at several non-interchangeable places, and the authors themselves acknowledge in Remark 2.3 that a weaker almost-additivity would suffice but is not treated. This is a scope limitation, not a flaw. I also searched for hidden circularity in the definition of fhom and the lower-bound proof. The definition of fhom uses the subadditive ergodic theorem, which is legitimate; the lower bound uses the cell formula only in the direction fhom(j) <= liminf F_epsilon(J_epsilon,A)/L^d(A) for competitors, which is established in Proposition 10.1 and does not assume the theorem being proved. The only mechanical defect is the rank bound in Remark 2.10/Lemma 8.8(4): the statement 'rank(dnu/d|nu|(x)) <= n-1' is imprecise without specifying the target space dimension (it should be n-1 for divergence-free measures valued in V ⊗ R^d with d = n, or more precisely rank <= dim V if d > dim V). However, this does not affect the main proof because the condition rank(j) <= n-1 is used only to define the correct strip dimension in Proposition 10.3, and the argument is insensitive to the exact constant. I therefore agree with the reader's ACCEPT verdict with moderate confidence. No further concern survives close reading; the honest output is a non-finding.","tokens_in":69068,"tokens_out":2071,"duration_ms":20380,"concrete_test":"Independently re-derive Lemma 6.6(iii) and the upper-bound gluing estimate (Eq. (9.27)) without invoking the full sigma-additivity of F3, replacing it with the almost-additivity condition suggested in Remark 2.3. If subadditivity and the cube-splitting energy estimates fail under the relaxed condition, then the F3 restriction is essential as stated; if they hold, the theorem could be extended and the F3 assumption is only a convenience. Additionally, verify the rank bound in Remark 2.10 against the continuous result [DeR16] for the specific space V ⊗ R^d used in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the manuscript in good faith and found no load-bearing defect in the central argument. Theorem 2.8 is a substantial Gamma-convergence result whose proof is explicit: the upper bound (Section 9) builds recovery sequences from the cell formula via Lemma 9.1, the gluing construction, and Proposition 7.4; the lower bound (Section 10) uses the blow-up method with Proposition 10.1 and Proposition 10.3 as the quantitative cores. The reader's weakest-assumption choice, sigma-additivity (F3), is genuinely the load-bearing structural condition: it is used in Lemma 6.6(iii) for subadditivity of the cell functional, in the upper-bound step Eq. (9.27) when splitting F_epsilon over cubes, and in the lower-bound estimates via additivity in Claims 1b/1c of Propositions 10.1 and 10.3. The authors explicitly flag the limitation in Remark 2.3, stating that almost-additivity would suffice but is not proved. This is an honest scope restriction, not an internal inconsistency. The only mechanical issue found is the rank bound in Remark 2.10 and Lemma 8.8(4), where the cited rank inequality for dnu/d|nu| should presumably be rank <= n-1 with n = dim V, or rank <= d in the ambient dimension; however, this remark is auxiliary and the main lower-bound argument only needs the rank condition in the form stated for the strip construction in Proposition 10.3, where j in V ⊗ R^d with rank(j) <= n-1 is used. Since the reader already identified this and it does not touch the central theorem, no verdict change is warranted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves an almost-sure Γ-convergence theorem, Theorem 2.8, for rescaled nonlinear minimum-cost flow functionals on stationary random graphs satisfying the geometric assumptions (G1)–(G3). For stationary random energies satisfying Lipschitz, linear-growth, and σ-additivity assumptions (F1)–(F3), and for prescribed discrete fluxes converging in the gKR topology, the rescaled functionals converge to a continuum divergence-constrained integral functional whose energy density is defined by a variational multi-cell formula built from uniform-flow operators. The proof combines deterministic discrete calculus, corrector estimates (Proposition 7.4), a blow-up argument in the style of Fonseca–Müller, and tangent-measure analysis of divergence measures. The upper bound is developed in Section 9 and the lower bound in Section 10, with quantitative non-asymptotic estimates in Propositions 10.1 and 10.3.","tokens_in":69408,"tokens_out":9804,"duration_ms":108300,"significance":"If the result stands, this is a substantial advance over previous periodic-graph homogenisation results: it removes periodicity, treats multi-species flows, allows nonconvex Lipschitz costs, and gives an explicit stationary effective density independent of the chosen uniform-flow operator. The paper is largely self-contained and contains several genuinely useful tools, notably the construction of uniform flows on stationary graphs (Proposition 5.4), the discrete corrector with quantitative gKR bounds (Proposition 7.4), and the strip-based singular lower bound (Proposition 10.3). There is no fitted parameter and no circularity: the effective density is defined by an independent cell formula and then shown to be the Γ-limit. The authors also honestly state the main structural limitation, the σ-additivity assumption F3, in Remark 2.3, where they note that only an almost-additive version would be needed for some applications but is not proved.","major_comments":[{"comment":"The rank condition is stated incorrectly and, taken literally, excludes the scalar case from the singular lower-bound argument. With n = dim V, the assertion rank(dν/d|ν|) ≤ n−1 in Remark 2.10 and Lemma 8.8(4) is false for scalar V: a non-zero singular density is a rank-one vector, so rank = 1 > 0 = n−1. The same hypothesis appears in Proposition 10.3, where it is used to define k = dim ker j. Since the singular part of the lower bound in Section 10.2 applies Proposition 10.3 to j = dξ/d|ξ|, the proof as written does not cover scalar flows, which are a central case of Theorem 2.8. The structural condition needed is that j has a non-trivial kernel, i.e. rank(j) ≤ d−1 (equivalently rank(j) ≤ min{n,d−1}), matching the wave cone of the divergence operator. Please correct the rank bound consistently in Remark 2.10, Lemma 8.8(4), and Proposition 10.3, and verify that the strip construction in Proposition 10.3 goes through with k = dim ker j in this corrected range.","section":"Remark 2.10, Lemma 8.8(4), Proposition 10.3"}],"minor_comments":[{"comment":"In the statement of the main theorem, µ is described as an element of M(U) at the end of the first paragraph, although the preceding sentence correctly places it in M(U;V); the target space should be M(U;V).","section":"Theorem 2.8"},{"comment":"The final sentence of the statement contains a duplicated and incomplete phrase: 'Finally, for L^d-a.e. x0 ∈ supp(|\\nu|) \\ E, we have For L^d-a.e. x0 ∈ R^d we have Tan_C(ν,x0) = {jL^d}'. This should be rewritten as a single sentence.","section":"Lemma 8.5"},{"comment":"The proof refers to 'the measure νxy in (10.2)' and to 'Remark 7.6', but (10.2) belongs to Section 10 and there is no Remark 7.6. These references should point to the corresponding construction in the proof of Proposition 7.6 or to an explicit equation in Section 7.","section":"Proposition 7.4, proof"},{"comment":"The measure νρ is defined twice in the same paragraph, once before the construction of the deformation Φρ and once after it; the first definition should be removed or the second definition labeled as the final one.","section":"Lemma 9.1, Step 1"},{"comment":"The sentence 'Lemma 6.8, implies that, as ε → 0, f_{ε/δ,R}(j0, Q1(x0/δ)) → f_hom(j0) vaguely' describes convergence of real numbers, so the adverb 'vaguely' is inappropriate and should be deleted.","section":"Section 10.1, 'Approximation 1'"}],"recommendation":"major_revision","confidential_remarks":"The rank-index error is the only substantive obstacle I see. In my view the central theorem is very likely correct and the fix is local, but the singular lower bound for scalar flows is not covered by the proof as written. I would be willing to recommend acceptance after the rank condition is corrected consistently and the strip argument is re-verified in the corrected range."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuine stochastic homogenisation result for nonlinear minimum-cost flow on stationary random graphs, and it should go to referees. The construction that makes it work—uniform-flow operators on graphs that are not periodic—is the real novelty, and the proof around it is careful and largely self-contained.\n\nWhat the paper actually does: it extends the periodic-graph results of Gladbach–Kopfer–Maas–Portinale to stationary random graphs satisfying (G1)–(G3), with nonconvex Lipschitz costs, linear growth, and multi-species flows. The limit is a continuum divergence-constrained variational problem with an effective density defined by a variational cell formula, not fitted. The upper bound (Section 9) and lower bound (Section 10) are both explicit; the blow-up machinery for the singular part (Proposition 10.3) is the heaviest piece and it is worked out in real detail. I checked the logical dependencies: no circularity, no hidden fitted parameters. The paper is honest about what it does not cover—Poisson point processes are excluded by (G1), and that is stated plainly.\n\nThe soft spots are proportional. The sigma-additivity assumption (F3) is the load-bearing structural hypothesis: it is used in the subadditivity of the cell formula (Lemma 6.6(iii)), the gluing step in the upper bound (eq. 9.27), and the additivity of the energy in the lower-bound corrections. The authors flag in Remark 2.3 that almost-additivity would suffice but is not proved; that is a real scope restriction, not a hidden flaw. The rank bound in Remark 2.10 and Lemma 8.8(4) indeed looks off—as stated, with n = dim V, it would make the singular part of a scalar divergence measure vacuous, which cannot be right. This is auxiliary and does not touch the main convergence proof, but it should be corrected before publication. The self-citations to [GK*23] are for published lemmas such as Lemma 7.2, which is standard reuse, not circularity.\n\nThe paper is written for specialists in calculus of variations, optimal transport, and stochastic homogenisation—readers who work with Gamma-convergence and blow-up methods will get the most out of it. The proof is long, but the main theorem is worth the effort.\n\nRecommendation: accept; send to knowledgeable referees. The core theorem is convincing and the uniform-flow construction is a reusable tool.","headline":"Strong stochastic homogenisation theorem with a genuinely new uniform-flow construction; the sigma-additivity assumption is a real limitation but the paper is honest about it and deserves refereeing.","tokens_in":69918,"tokens_out":4787,"would_cite":true,"duration_ms":54331,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35B27","49J45","49Q22","05C21"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that nonlinear minimum-cost flow problems on stationary random graphs converge, after rescaling, to a continuum divergence-constrained variational problem with an explicit effective energy density.","keywords":["stochastic homogenisation","minimum-cost flow","Gamma-convergence","random graphs","discrete-to-continuum limits","optimal transport","div-quasiconvexity","multi-species flows"],"falsifier":"Compute the cell energies $f_{\\varepsilon,R}(j,Q)$ on a stationary random graph satisfying (G1)–(G3) for two different admissible uniform-flow operators $R$; if the limits differ for some $j$, the claimed independence of the effective density from $R$ fails. More directly, simulate the rescaled minimum-cost flow problem on a random Voronoi tessellation with $m_\\varepsilon$ approaching a smooth divergence field and compare the limiting minimal energy with the integral of $f_{\\mathrm{hom}}$; a mismatch contradicts the $\\Gamma$-convergence statement.","tokens_in":68869,"feed_emoji":"🌐","tokens_out":11275,"duration_ms":105167,"temperature":0.7,"pith_summary":"Nonlinear minimum-cost flow problems ask for the cheapest way to route prescribed amounts of mass through a graph, with possibly nonconvex costs on edges. The paper's main theorem states that, for a large class of stationary random graphs in $\\mathbb{R}^d$, these problems have a well-defined large-scale limit: after rescaling, the discrete functionals $\\Gamma$-converge almost surely to a continuum functional in which a random effective cost density is minimised among vector fields with prescribed divergence. The effective density is built from a variational cell formula, and its construction is the main technical achievement because the graph has no periodic structure. A consequence is convergence of the minimal costs and of approximate minimisers, covering multi-species flows and the Wasserstein distances induced by random edge lengths.","feed_headline":"Random-graph minimum-cost flows homogenise to a continuum problem","feed_subtitle":"Rescaled flow costs converge almost surely to a variational limit with an explicit effective density","key_machinery":"The central objects are the discrete uniform-flow operator $R$ — a bounded linear map sending each constant tensor $j$ to a divergence-free discrete field $Rj$ whose embedded rescaling converges to $j\\,\\mathcal{L}^d$ — and the associated cell formula $f_{\\omega,\\varepsilon,R}(j,A)$, defined as the minimal energy of a divergence-free discrete field matching $R_\\varepsilon j$ near $\\partial A$. The homogenised density $f_{\\omega,\\mathrm{hom}}(j)$ is the $\\varepsilon\\to 0$ limit of the cell energy per unit volume, shown to exist by subadditivity and the subadditive ergodic theorem and to be independent of $R$. The lower bound is carried by correctors for the discrete divergence equation and by the blow-up of divergence measures: the measure differentiation theorem shows that local rescalings of the flux converge to tangent measures that are constant-density and divergence-free, so the cellular inequality $f_{\\omega,\\mathrm{hom}}(j_0) \\le$ local energy density can be tested on near-representatives.","core_discovery":"Under the paper's standing assumptions — a stationary random graph satisfying (G1)–(G3) (no large gaps, quantitative connectedness, bounded edge lengths) and a stationary random energy satisfying (F1)–(F3) (Lipschitz, linear growth, $\\sigma$-additivity) — Theorem 2.8 asserts that for any sequence of flux data $m_\\varepsilon$ converging to $\\mu$ in the Kantorovich–Rubinstein norm, the constrained flow functionals $F_{\\omega,\\varepsilon}(\\cdot|m_\\varepsilon)$ $\\Gamma$-converge almost surely to $F_{\\omega,\\mathrm{hom}}(\\cdot|\\mu)$. The limit is an integral functional over vector-valued measures with a density $f_{\\omega,\\mathrm{hom}}$ that is stationary, lower semicontinuous, div-quasiconvex and of linear growth, and the minimal values converge to the minimal value of the limit. The effective density is obtained as $f_{\\omega,\\mathrm{hom}}(j) = \\lim_{\\varepsilon\\to 0} f_{\\omega,\\varepsilon,R}(j,Q)/\\mathcal{L}^d(Q)$, where $f_{\\omega,\\varepsilon,R}$ is the infimum of the energy over divergence-free discrete fields that coincide with a chosen uniform-flow operator $R$ near the boundary of the unit cube; the result shows the limit is independent of $R$. The proof combines existence of uniform flows on graphs satisfying (G1)–(G2), correctors for the discrete divergence equation, and a blow-up lower bound using tangent measures of divergence measures.","pith_inferences":["The paper's Remark 2.3 suggests that $\\sigma$-additivity (F3) can be relaxed to almost-additivity for well-separated sets; once proved, the method should extend to mildly nonlocal graph energies whose nonlocality fades in the large-scale limit.","Applied to iid edge weights on $\\mathbb{Z}^d$ with homogeneous linear costs, the result implies convergence of rescaled graph Wasserstein distances to a deterministic norm; identifying that norm with the first-passage percolation time constant would link the homogenisation limit to the shape theorem, but that identification is not proved here.","A natural next step is quantitative homogenisation: for graphs with finite range of dependence, rates for the cell formula could be obtained from concentration or variance bounds on the subadditive limit, going beyond the paper's almost-sure statement."],"forward_implications":["When costs are 1-homogeneous edge lengths, the rescaled graph 1-Wasserstein distances converge almost surely to a homogenised distance induced by the norm $f_{\\omega,\\mathrm{hom}}$; under ergodicity the limiting norm is deterministic.","The $\\Gamma$-convergence and compactness imply convergence of minimal costs and of approximate minimisers: any cluster point of embedded discrete minimisers minimises the continuum functional, and a unique continuum minimiser attracts the discrete ones.","Multi-species flows, where measures take values in a finite-dimensional vector space, are covered, so the continuum limit applies to multi-commodity transportation on random networks.","The limiting energy density is automatically div-quasiconvex and of linear growth, matching the expected structure of variational limits under divergence constraints."],"supporting_citations":[{"why":"Blow-up method for lower semicontinuity in L1, adapted here to the discrete lower-bound argument.","marker":"[FoM92]"},{"why":"Provides div-quasiconvexity and lower semicontinuity for the continuum limit functional.","marker":"[FoM99]"},{"why":"Subadditive ergodic theorem used to define the homogenised density as an almost-sure limit.","marker":"[Kin73]"},{"why":"Global-local subadditive ergodic theorem used for the cell-ratio limit defining f_hom.","marker":"[LiM02]"},{"why":"Earlier periodic homogenisation of dynamical optimal transport; supplies discrete calculus and corrector tools generalised here.","marker":"[GK∗23]"},{"why":"Stochastic homogenisation on random lattices with graph assumptions similar to (G1)–(G3), used as a model for the discrete setting.","marker":"[ACG11]"},{"why":"Rank bound for divergence measures, used to control the singular part in the lower bound.","marker":"[DeR16]"},{"why":"Vectorial Kantorovich–Rubinstein duality used in the corrector construction.","marker":"[Cio21]"},{"why":"Measure differentiation, Lebesgue points and tangent measures used in the blow-up argument.","marker":"[AFP00]"}],"fun_headline_variants":["Random graphs: min-cost flows homogenise to continuum","Gamma-limit for random-graph flow costs","Min-cost flows on random graphs get a continuum limit","Stochastic homogenisation yields effective flow cost on random graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The energy must be exactly additive over disjoint regions (condition F3); the cell formula, the subadditivity of the cell energies, and the upper-bound gluing argument all use this, and the paper only notes that a milder almost-additivity condition would probably suffice without proving it.","fun_headline_variants_meta":{"raw":{"variants":["Random graphs: min-cost flows homogenise to continuum","Gamma-limit for random-graph flow costs","Min-cost flows on random graphs get a continuum limit","Stochastic homogenisation yields effective flow cost on random graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002033,"raw_usage":{"total_tokens":7945,"prompt_tokens":990,"completion_tokens":6955,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":6892}},"tokens_in":606,"tokens_out":6955,"duration_ms":48565,"temperature":1.0,"reasoning_tokens":6892,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:50:59.095700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the cell energies $f_{\\varepsilon,R}(j,Q)$ on a stationary random graph satisfying (G1)–(G3) for two different admissible uniform-flow operators $R$; if the limits differ for some $j$, the claimed independence of the effective density from $R$ fails. More directly, simulate the rescaled minimum-cost flow problem on a random Voronoi tessellation with $m_\\varepsilon$ approaching a smooth divergence field and compare the limiting minimal energy with the integral of $f_{\\mathrm{hom}}$; a mismatch contradicts the $\\Gamma$-convergence statement.","supporting_citations":[],"review_version":1}